I Only Wanted a Vector. Then I Fell Through Nine Layers.
A small embedding service for a voice assistant sent me down through tokens, runtimes, kernel libraries and CPU lanes. The biggest speed-up came from a trap in the model file, not from a faster language.